TALOS NIO

TALOS is an offline incremental training and evaluation stack for Nymeria dual-IMU data.

The current pipeline combines:

• 100 Hz ESKF propagation on the primary IMU
• a spectral MLP that predicts mean local velocity and log-variance
• physics-based guardrails such as LAID, ZARU, CAU, NPP tracking, and a positional cage

Core Files

• incremental_train.py - main training and physical ESKF evaluation loop
• SMLP.py - spectral neural model used by the trainer
• nymeria_loader.py - Nymeria loading, alignment, windowing, and augmentation
• bulwark.py - per-axis local-velocity clamp
• laid.py - lever-arm differential consistency checks
• npp.py - neck pivot point tracking
• halo.py - orientation cage observer, currently instantiated but disabled in evaluation
• telemetry.py - telemetry CSV and diagnostic plots
• reporting.py - ntfy / Notion publishing helpers
• darwin.py - mutation-based recovery for stagnant runs
• cache_builder.py - cache generation utility
• plot_shelby.py - Shelby trajectory plotting helper
• scan_dataset.py - dataset inspection utility

Runtime Notes

• incremental_train.py is CUDA-only in the current repository state.
• The model is trained with AdamW and ReduceLROnPlateau.
• trans labels are mean local velocity over each window; quat labels are carried through the dataset but are not used by the current loss.
• bulwark() clips implausible predictions; it does not zero them.
• HALO orientation clamping, LAID yaw anchor, and the dynamic observation covariance path are present but disabled by default in the evaluation loop.

Quick Start

python cache_builder.py
python incremental_train.py
python plot_shelby.py

If you only need the detailed technical state, read TALOS.md.

Status

Active R&D / prototype codebase under rapid iteration.